DEV Community

雷文
雷文

Posted on

I Built DeepRead: An Evidence-First Reading Skill for Articles, Books, and PDFs

#ai #showdev #opensource #productivity

I recently built an open-source project called DeepRead to solve a problem I kept seeing with AI reading tools.

Most tools can summarize a document quickly, but they often blur the line between the author's actual claims, the evidence provided by the source, and the model's own inferences. With long documents, source locations and the context around numerical claims can also disappear.

DeepRead turns articles, books, PDFs, and document sets into structured reports organized around claims, evidence, data, and relationships. Important claims are linked to supporting evidence and source locations. If the source does not provide evidence, the report says so instead of filling the gap.

Five Reading Modes

  • quick — Get the main idea and key takeaways.
  • deep — Analyze the core claim and argument structure.
  • map — Generate a knowledge map, confidence levels, evidence pairs, data tables, relationships, and recall questions.
  • feynman — Explain the material in plain language, identify knowledge gaps, correct them against the source, and create a review plan.
  • book — Process a whole book chapter by chapter.

DeepRead can also compare 2–10 documents and identify agreements, conflicts, and complementary evidence.

Two Ways to Use It

Agent Skill for Codex and Claude Code

The portable skill has no additional runtime dependencies:

npx skills@latest add xiehuan123/dsh-deepread
Enter fullscreen mode Exit fullscreen mode

Example prompt:

Deep-read architecture.pdf in knowledge-map mode.
For every important claim, show the supporting evidence and page location.
Enter fullscreen mode Exit fullscreen mode

Full DeepSeek Harness Plugin

The full plugin includes a reading panel, PDF extraction, background jobs, live progress, batch comparison, cost estimates, and Markdown, HTML, or XMind-compatible export.

dsh plugin --profile web add dsh-deepread
Enter fullscreen mode Exit fullscreen mode

Try It

GitHub:

https://github.com/xiehuan123/dsh-deepread

v0.5.4 Release:

https://github.com/xiehuan123/dsh-deepread/releases/tag/v0.5.4

The repository contains three real output examples rather than hand-written mockups. DeepRead has already been included in several DeepSeek Harness and DSH community collections. Its GitHub Awesome Copilot submission has also passed the automated specification, lint, installation, and version checks and is waiting for maintainer review.

Known Limitations

  • Scan-only PDFs need OCR first.
  • Some websites with aggressive anti-bot protection may require you to paste the article text directly.
  • Evidence-heavy reports use more tokens than conventional summaries.

The project is MIT licensed. I would especially like feedback on one question:

Is explicit claim-to-evidence tracing useful in your reading workflow, or does it make the report too heavy?

Issues and pull requests are welcome.

Suggested DEV tags: showdev, opensource, ai, productivity

Suggested Show HN title: Show HN: DeepRead – Evidence-first reading for books, articles, and PDFs

Top comments (0)